PhD (M/F) on deep learning and applications using hyperspectral images and multimodal data

Job title:

PhD (M/F) on deep learning and applications using hyperspectral images and multimodal data


Job description

Offer DescriptionThe recruited person will be charged to IRIT at Paul Sabatier University and will have the opportunity to perform missions at the University of Bucharest. The Institute of Research in Computer Science of Toulouse (IRIT), one of the largest Joint Research Units (UMR 5505) at the national level, is a cornerstone of research in Occitania with its 600 members and about a hundred external collaborators. Due to its multi-institutional nature (CNRS, Toulouse universities), its scientific impact, and its interactions with other fields, the laboratory constitutes one of the structuring forces in the landscape of computer science and its applications in the digital world. Through its cutting-edge work and dynamic approach, our unit has defined its identity and gained undeniable visibility, positioning itself at the heart of the evolution of local structures: the University of Toulouse, as well as various initiatives from future investment programs (LabEx CIMI, IRT Saint-Exupéry, SAT TTT, etc.).
The “Information Systems” team is one of the largest teams in the laboratory with 20 teacher-researchers. The research focuses on data, which is at the core of modern information systems. The data is massive (“Big Data”), produced by humans or systems (satellite systems, social networks, medical imaging, sensors, video surveillance systems). The research aims to design and develop methods, models, languages, algorithms, and software tools that allow simple and efficient access to relevant information to improve its use, facilitate analysis, and aid decision-making. Our research covers the entire data processing chain, from raw data to refined data accessible to users seeking information, wishing to visualize it, and perform decision-making, exploratory, and predictive analyses.
Josiane Mothe ( ) has been a professor since 2002.With a history of over 150 years and a constantly reconfirmed reputation, the University of Bucharest is today a dynamic and inclusive academic space, characterized by creativity, innovation, and pragmatism. With over 34,000 students, 1,300 professors and 600 researchers, 1,200 employees as administrative staff, the University of Bucharest is a strong and supportive community, constantly working to ensure the highest quality of provided services, guaranteeing its graduates easy and effective assimilation in the labor market, regardless of the field or program of study. The University of Bucharest occupies the first position among Romanian universities, and in the field of Computer Science, it is ranked between 551-600 in the QS World University Rankings by Subject.The AI Lab at the University of Bucharest, led by Prof. Radu Tudor Ionescu ( ), conducts fundamental and applied research in artificial intelligence, machine learning, deep learning, computer vision, image processing, text mining, computational linguistics, medical imaging, and signal processing. Its members regularly publish in top journals (TPAMI, IJCV) and conferences (CVPR, NeurIPS, ICCV, ECCV, ACL, ACMMM, EMNLP, NAACL, WACV, ECML-PKDD, INTERSPEECH, EACL) in the field.Hyperspectral imaging, used in medicine and agriculture, is a valuable tool for monitoring the human body and the earth’s surface. However, image processing methods recently developed for color images have had limited success when applied to hyperspectral images, due to the challenges associated with collecting and annotating hyperspectral data in large quantities. To overcome this challenge, our work aims to develop a novel self-supervised and multimodal learning techniques for hyperspectral images, as well as neural architectures exploiting unlabeled multimodal data. This work fosters collaboration and knowledge exchange between the fields of computer science, medicine and agriculture.In both fields, medical and spatial imaging, data are very voluminous and abundant. However, their annotation is very costly, as it requires experts who are few in number and who have other priorities; this is particularly true in the medical field.
The aim of this thesis is to study models that take advantage of the abundance of non-annotated data. Not only images are key for these applications, but other forms of data such as structured data and texts can also be combined for better domain understanding.
Four research questions will structure this work:RQ1: Combining supervised and unsupervised learning
In the field of machine translation for low-resource languages, the combination of large language models trained on a highly resource language (such as English, for example) combined with low-resource language learning by considering equivalences between pairs of sentences in different languages has shown its effectiveness. We believe that learning the internal structures of images on non-annotated data, combined with learning on sparse annotated data is a promising avenue. For example, the combination of convolutional neural networks with transformer models in a semi-supervised way is promising.RQ2: Self-supervised learning
Self-supervised methods can exploit the abundance of unlabeled data available. The aim is to draw on the inherent structure or context of the data to generate signals for supervision. Moreover, working on medical data and Earth observation data could help develop more robust representations.RQ3: Combining data in deep neural network architectures
On Earth observation images, spectrum combinations have been used to extract indices that can then be used to detect phenomena, such as the vegetation indices used to detect weeds. We will study the adaptation of this principle to the medical field, for data classification and calibration. Transposing the principles of using engineered indices from different image spectra in the medical field is an open question.RQ4: Combining different forms of data
Multimodal data can be useful in many applications. How could these data be combined is an open question. In this PhD we will work on new models that can combine text, images and data. This combination could be start from the embedding layer of the deep models, or could be based on some new forms of attention, or the fusion can happen as a late process. The impact of the architectural choices will be studied. LLMs will be used.Where to apply WebsiteRequirementsResearch Field Computer science Education Level PhD or equivalentResearch Field Mathematics Education Level PhD or equivalentLanguages FRENCH Level BasicResearch Field Computer science Years of Research Experience NoneResearch Field Mathematics » Algorithms Years of Research Experience NoneAdditional InformationAdditional commentsEducational background and Skills :
– A Master’s degree (or equivalent) in Computer Science or Data Science,
– Knowledge of (theory / practical project,..) : Hyperspectral imaging, Machine Learning, Deep Learning including convolutional neural networks (CNNs) and transformer models, self-supervised learning methods, Large language Models, Multimodal Learning.
– Programming: Proficiency in programming languages commonly used in machine learning (e.g., Python, TensorFlow, PyTorch).
– Demonstrated interest in research through previous projects, internships, or publications.Application Requirements:
– A detailed CV highlighting relevant experience and skills (one or several as detailed in te description).
– A cover letter explaining the candidate’s interest in the position and relevant qualifications.
– Academic transcripts from previous degrees.
– Contact information for at least two academic or professional references.
– Samples of previous research work or publications.
– Candidates can be first year PhD students in Romania Website for additional job detailsWork Location(s)Number of offers available 1 Company/Institute Institut de Recherche en Informatique de Toulouse Country France City TOULOUSE GeofieldContact CityTOULOUSE WebsiteSTATUS: EXPIREDShare this page

Expected salary



Job date

Sat, 08 Jun 2024 03:20:33 GMT

To help us track our recruitment effort, please indicate in your email/cover letter where ( you saw this job posting.

To apply for this job please visit

Job Location